You say what now?
“The blue softly engine dreamt under ranly trees of sparkled thoughtness in a slow cloud sleep.”
No, that is not me losing it after watching an endless flow of matrices while studying the nitty-gritty of large language models, or LLMs for short. These days, when people say ‘AI,’ they usually mean large language models that underpin tools like ChatGPT or Midjourney. To understand these tools is to understand LLMs. Annoyingly, that is easier said than done. It is a pretty deep rabbit hole, which quickly dissolves into a cacophony of matrices, attention algorithms and neural networks. I still need to figure out how to write it as blog posts. And as they say: “If you cannot explain it, you don’t really understand it”. So, for now, I’ll keep working on understanding it, and we’ll keep the LLM as a black box; we only add that we can “train” it. The more we “train” an LLM, the better it will respond, and the better the tools built with it are.
The quote you saw is output from a novice LLM that has just started its training. It has dreamed up words such as “ranly” and “thoughtness”. The sentence is largely gibberish.
An LLM that has advanced further in its training might say: “The blue engine dreamed under tall trees and thought about the slow clouds”. The words are correct, but the sentence makes very little sense, except perhaps in a sci-fi setting where engines dream about clouds of various speeds. The LLM is trying to say: “The blue locomotive slowed beneath the pine canopy, its engine sighing as clouds drifted across the fading sky”, which it can do with outstanding efficiency when it completes its training.
The billion-pronged tuning fork
When I was about ten years old, we learned to play the recorder (the duct flute) in school. I must have driven my family close to madness with the discordance of harsh squeaks, which, I am told, penetrated the entire house even behind my room’s closed door. But lo-and-behold, a few months later, I played “Edelweiss” at our school’s Christmas party. Now, it was not a world-class presentation of the song, mind you, but I think the audience could identify the tune. The “Edelweiss” was my locomotive—although it never ran flawlessly, at least I got it moving.
Like me, the bot learns too. It’s just that an LLM’s sound is the human language, and its ear is mathematics: the only language computers truly understand. In the simplest sense, its training is just practice, but at an impossible scale. An LLM reads through billions of examples of text during its training: articles, code, conversations and song lyrics. We can argue if it should, but it does. While reading, it plays a game of prediction: it looks at part of a sentence and tries to guess what comes next. If the guess is wrong—and early on, it almost always is—it tinkers with its own parameters, its configuration, before trying again. After trillions of modifications, it sounds like a human, and you can talk to ChatGPT.
The computer doesn’t understand words in the way we do. It doesn’t know what a “tree” or a “dream” is. But by seeing how words appear together, which ones are likely to follow which, it builds an incredibly detailed map of how language behaves. The map tells it that “coffee” is often near “cup” and rarely near “planet”. When you later ask it to write a poem or answer a question, it isn’t recalling memorized text; it’s following the contours of that learned map to predict what a good continuation should look like. It tries to guess what a human would say in its place. All this happens without any real awareness. The model doesn’t know it’s learning. It is simply repeating a process of prediction and correction, again and again, until the results sound human.
I swear, at least I think I do.
ChatGPT will always give you an answer. It never replies: “Umm, not sure”. This is baked into how LLMs work. They are prediction engines, and that’s all they are. An LLM always fulfills its function: guessing how the text should continue. It will do that even if the guess is bad. For it, guessing is its core mission, and failure is not an option. That is why we end up in alternative realities where engines dream of clouds. This is so common that we have given it a word of its own. When we say an AI “hallucinates,” we are borrowing from the human world. Our own brains do something similar when perception or memory falters. A witness may recall a car’s color or a face that was never there—a false memory. People with brain injuries sometimes confabulate, inventing missing details so their story stays coherent. The visually impaired may experience Charles Bonnet hallucinations, seeing clear scenes their eyes no longer deliver. Even healthy minds fill gaps: we see faces in clouds, or feel déjà vu in places we have never been. In all these cases, the brain favors a story that makes sense over admitting uncertainty. An LLM will do the same. When it lacks facts, it confidently predicts what “sounds right”, describing a scene it never actually saw.
Previous articles in this series
- Part 1: The Philosophy of Intelligence